Automated Left Ventricle Ischemic Scar Detection in CT Using Deep Neural Networks

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Identifying cardiac scar tissue is crucial for patient diagnosis and intervention planning.
• Late gadolinium enhancement (LGE) magnetic resonance imaging (MRI) is the gold standard but has contraindications.
• Computed tomography angiography (CTA) is a faster alternative with fewer contraindications but limited scar imaging capability.

Purpose of the Study:

• To develop a scar detection method for routine CTA imaging using deep convolutional neural networks (CNNs).
• To ensure the method relies solely on anatomical information and integrates with existing clinical workflows.
• To provide an alternative scar imaging technique when MRI is contraindicated.

Main Methods:

• A CNN was trained and validated using LGE MRI data from 200 patients (83 with scar), with segmentation masks as input.
• 3D left ventricle meshes were generated from MRIs, and anatomical masks with LGE scar labels were extracted.
• The trained CNN was tested on an independent CTA dataset (25 patients) using automated segmentation, and performance was compared to manual expert readings.

Main Results:

• The CNN achieved 84.7% cross-validated accuracy (AUC: 0.896) for scar detection in MRI data.
• On independent CTA data, the network achieved 88.3% accuracy (AUC: 0.901) without further training.
• The automated pipeline demonstrated superior performance compared to manual interpretation by clinicians.

Conclusions:

• Automatic ischemic scar detection is feasible using routine cardiac CTA without scar-specific imaging or contrast agents.
• This method requires a single cardiac cycle acquisition.
• Clinical implementation offers near-zero additional cost for scar detection, image triage, reduced reading times, and guided clinical decisions.